Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-19 are rejected under 35 U.S.C. § 101 because the claimed invention is directed to an abstract idea without significantly more.
Independent Claims
Step 2A Prong One: Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes, independent claim 1, for example, recites an abstract idea in the form of mental processes. A mental process is a process that “can be performed in the human mind, or by a human using a pen and paper” (MPEP § 2106.04(a)(2)(III), paragraph 1). Examples of mental processes include “observations, evaluations, judgments, and opinions” (MPEP § 2106.04(a)(2)(III), paragraph 2).
The following limitations of claim 1 are mental processes:
“search for proposed molecules” [This limitation is a mental process that can be performed by observation, evaluation, judgment, and opinion, since the claim merely recites searching at a high degree of generality, without any specific methodology that precludes this limitation from being a mental process.]
“predict a likelihood of the desired property being present in the proposed molecules” [This limitation is a mental process that can be performed by observation, evaluation, judgment, and opinion, since the claim merely recites prediction of a desired result at a high degree of generality, without any specific methodology that precludes this limitation from being a mental process.]
“approximate a degree of similarity between the proposed molecules and molecules known to possess the desired property, wherein the search is guided by the […] predictions and the […] approximations.” [This limitation is a mental process that can be performed by observation, evaluation, judgment, and opinion, since the claim merely recites approximating and searching a degree of similarity at a high degree of generality, without any specific methodology that precludes this limitation from being a mental process.]
The following limitations of claim 10 are mental processes:
“predict a likelihood that each respective one of a plurality of proposed molecules has the desired property” [This limitation is a mental process that can be performed by observation, evaluation, judgment, and opinion, since the claim merely recites prediction of a desired result at a high degree of generality, without any specific methodology that precludes this limitation from being a mental process.]
“approximate a statistical similarity between each respective one of the proposed molecules and a set of molecules represented in a stored dataset of training molecules, wherein the dataset of training molecules contains data regarding a set of molecules known to possess the desired property” [This limitation is a mental process that can be performed by observation, evaluation, judgment, and opinion, since the claim merely recites approximating a degree of similarity at a high degree of generality, without any specific methodology that precludes this limitation from being a mental process.]
“conduct a search through a space of candidate molecules to propose molecules likely to have the desired property, wherein the search is guided by the […] predictions and by the […] approximations.” [This limitation is a mental process that can be performed by observation, evaluation, judgment, and opinion, since the claim merely recites searching at a high degree of generality, without any specific methodology that precludes this limitation from being a mental process.]
Independent claim 19 recites limitations that are analogous to those of claim 10. Therefore, the above analysis for claim 10 is applied to claim 19.
Step 2A Prong Two: Does the claim recite additional elements that integrate the judicial exception into a practical application?
No. The judicial exception recited in the above discussed claims is not integrated into a practical application.
Independent claims 1, 10, and 19 recite the following additional elements, but these additional elements are not sufficient to integrate the judicial exception into a practical application:
“A computer system for designing molecules with a desired property, the computer system comprising: a computer-based processor; and a computer-readable medium storing computer-readable instructions that, when executed by the computer-based processor, cause the computer-based processor to embody” (claim 1), “A computer-based method for designing molecules having a desired property, the computer-based method comprising: providing a computer system comprising: a computer-based processor; and a computer-readable medium storing computer-readable instructions that, when executed by the computer-based processor, cause the computer-based processor to embody” (claim 10), and “A non-transitory computer readable medium having stored thereon computer-readable instructions that, when executed by a computer-based processor, cause the computer-based processor to embody” (claim 19) [These elements constitute no more than mere instructions to apply the judicial exception using generic computer components (MPEP § 2106.04(d)(I)). These additional elements merely invoke the use of generic computer components, such as a processor or compute readable medium, as tools to perform the abstract idea, and do not place any limitations on the abstract idea other than the use of such generic computer components. Therefore, these additional elements do not integrate the judicial exception into a practical application.]
“a computer-based search engine”, “a computer-based property predictor”, and “a computer-based energy model” (claims 1, 10, and 1) [These elements are regarded as software-implemented computer functions that perform the mental processes described above. While the claim uses terms like “search engine” and “predictor,” the structure of these elements are claimed only generically as computer-implement fucntions. Therefore, these elements constitute no more than mere instructions to apply the judicial exception using generic computer functions (MPEP § 2106.04(d)(I)) as tools to perform mental processes. Therefore, these additional elements do not integrate the judicial exception into a practical application.]
Step 2B: Does the claim recite additional elements that amount to significantly more than the judicial exception?
No. The claims do not include additional elements that are sufficient for the claims to amount to significantly more than the judicial exception.
Additional elements that are mere instructions to apply an exception do not constitute significantly more than a judicial exception under MPEP § 2106.05(I)(A). Therefore, those additional elements identified above in the Prong One analysis as mere instructions to apply an exception do not constitute significantly more.
Dependent Claims
The remaining dependent claims being rejected do not recite additional elements, whether considered individually or in combination, that are sufficient to integrate the judicial exception into a practical application or amount to significantly more than the judicial exception. The limitations of the dependent claims are analyzed as follows.
Dependent Claim 2
“a dataset of training molecules in the computer-readable medium, wherein the dataset of training molecules contains data regarding the molecules known to possess the desired property” [This element constitutes no more than mere instructions to apply the judicial exception using generic computer components (MPEP § 2106.04(d)(I)), namely the use of computer-readable medium to store data.]
Dependent Claim 3
“wherein the computer-based property predictor has been trained, in a supervised manner, to predict the likelihood of the desirable property being present in the proposed molecules, using the dataset of training molecules.” [These elements are additional elements besides the abstract idea, but they constitute no more than mere instructions to apply the judicial exception using generic computer functions (MPEP § 2106.04(d)(I)), namely the generic computer function of machine learning. These additional elements merely invoke the use of generic machine learning as a tool to apply an abstract idea, namely as a tool to perform the mental processes of prediction associated with the predictor.]
Dependent Claim 4
“wherein the computer-based energy model has been trained, in an unsupervised manner, to model a distribution of data in the dataset of training molecules.” [These elements are additional elements besides the abstract idea, but they constitute no more than mere instructions to apply the judicial exception using generic computer functions (MPEP § 2106.04(d)(I)), namely the generic computer function of machine learning. These additional elements merely invoke the use of generic machine learning as a tool to apply an abstract idea, namely as a tool to perform the mental processes of approximation associated with the energy model.]
Dependent Claim 5
“wherein, for each molecule that the computer-based search engine proposes: […] predicts a likelihood that the proposed molecule has the desired property; and the […] approximates a degree to which the proposed molecule is similar to molecules represented in the dataset of training molecules.” [These elements are mental processes that can be performed by observation, evaluation, judgment, and opinion for the same reasons given for the analogous limitations in the parent independent claim. That is, these recitations merely recite prediction or approximation at a high degree of generality, without any specific algorithm or technical methodology.]
The limitations of “computer-based property predictor” and “computer-based energy model” are given the same analysis as in the independent claims.
Dependent Claim 6
“a dataset of proposed molecules in the computer-readable medium” [This element constitutes no more than mere instructions to apply the judicial exception using generic computer components (MPEP § 2106.04(d)(I)), namely the use of computer-readable medium to store data.]
“wherein, for each molecule proposed by the search engine, the system determines whether to store data associated with that proposed molecule in the dataset of proposed molecules, based on the predicted likelihood that that proposed molecule has the desired property, and based on the approximate degree to which that proposed molecule is similar to the molecules represented in the dataset of training molecules.” [These elements are mental processes that can be performed by observation, evaluation, judgment, and opinion.]
Dependent Claim 7
“wherein the search engine scores each respective one of the proposed molecules based on the predicted likelihood of that proposed molecule having the desired property, and based on the approximate degree to which that proposed molecules is similar to the molecules represented in the dataset of training molecules” [These elements are mental processes that can be performed by observation, evaluation, judgment, and opinion. While the claim recites that the scoring is based on certain elements, the term “based on” does not require any specific technical methodology that is distinguished from a mental process.]
“storing the data associated with that proposed molecule in the dataset of proposed molecules if the score for that proposed molecule is above a threshold score or higher than assigned composite scores for other proposed molecules.” [These elements are considered to be additional elements besides the abstract idea. However, these elements are merely “adding insignificant extra-solution activity to the judicial exception” (MPEP § 2106.05(g)) since they merely amount to necessary data gathering or outputting, which identifies is identified in MPEP § 2106.05(g) as a form of extra-solution activity. Furthermore, for purposes of Step 2B analysis, these elements are well‐understood, routine, and conventional because they are limitations of “storing and retrieving information in memory,” which MPEP § 2106.05(d)(II) identifies as well‐understood, routine, and conventional computer functions.]
Dependent Claim 8
“wherein the computer-based property predictor is a computer-based graph neural network (GNN), the energy model is a Deep Energy Estimator Network (“DEEN”), and the search engine executes Monte Carlo tree searching (MCTS).” [These elements are considered to be additional elements besides the abstract idea but they do no more than generally link the use of a judicial exception to a particular technological environment or field of use (MPEP § 2106.05(h)), namely the technology environments of GNN, DEEN, and MCTS.]
Dependent Claim 9
“wherein the desired property is inhibition of SARS-CoV-1.” [This element merely further defines the mental processes recited in the independent claim, and does not introduce any further additional elements besides the abstract idea.]
Dependent Claim 11
“storing the dataset of training molecules in the computer-readable medium” [These elements are considered to be additional elements besides the abstract idea. However, these elements are merely “adding insignificant extra-solution activity to the judicial exception” (MPEP § 2106.05(g)) since they merely amount to necessary data gathering or outputting, which identifies is identified in MPEP § 2106.05(g) as a form of extra-solution activity. Furthermore, for purposes of Step 2B analysis, these elements are well‐understood, routine, and conventional because they are limitations of “storing and retrieving information in memory,” which MPEP § 2106.05(d)(II) identifies as well‐understood, routine, and conventional computer functions.]
“training, in a supervised manner, the computer-based property predictor to predict a likelihood that the desirable property is present in a molecule using the dataset of training molecules.” [These elements are additional elements besides the abstract idea, but they constitute no more than mere instructions to apply the judicial exception using generic computer functions (MPEP § 2106.04(d)(I)), namely the generic computer function of machine learning. These additional elements merely invoke the use of generic machine learning as a tool to apply an abstract idea, namely as a tool to perform the mental processes of prediction associated with the predictor.]
Dependent Claim 12
“training, in an unsupervised manner, the energy model to model a distribution of data regarding the training molecules in the dataset of training molecules.” [These elements are additional elements besides the abstract idea, but they constitute no more than mere instructions to apply the judicial exception using generic computer functions (MPEP § 2106.04(d)(I)), namely the generic computer function of machine learning. These additional elements merely invoke the use of generic machine learning as a tool to apply an abstract idea, namely as a tool to perform the mental processes of approximation associated with the energy model.]
Dependent Claim 13
“conducting the search through the space of candidate molecules with the computer-based search engine guided by the property predictor's predictions and the energy model's approximations.” [These elements are mental processes that can be performed by observation, evaluation, judgment, and opinion for the same reasons given for the analogous limitations in the parent independent claim. That is, these recitations merely recite prediction or approximation at a high degree of generality, without any specific algorithm or technical methodology.]
Dependent Claim 14
“predicting, with the computer-based property predictor, a likelihood that that proposed molecule has the desired property; and approximating, with the computer-based energy model, a degree to which that proposed molecule is similar to molecules represented in the dataset of training molecules.” [These elements are mental processes that can be performed by observation, evaluation, judgment, and opinion for the same reasons given for the analogous limitations in the parent independent claim. That is, these recitations merely recite prediction or approximation at a high degree of generality, without any specific algorithm or technical methodology.]
Dependent Claim 15
“storing, in the computer-readable medium, a dataset of proposed molecules” [These elements are considered to be additional elements besides the abstract idea. However, these elements are merely “adding insignificant extra-solution activity to the judicial exception” (MPEP § 2106.05(g)) since they merely amount to necessary data gathering or outputting, which identifies is identified in MPEP § 2106.05(g) as a form of extra-solution activity. Furthermore, for purposes of Step 2B analysis, these elements are well‐understood, routine, and conventional because they are limitations of “storing and retrieving information in memory,” which MPEP § 2106.05(d)(II) identifies as well‐understood, routine, and conventional computer functions.]
“for each molecule that the computer-based search engine proposes: determining whether to store data associated with that proposed molecule in the dataset of proposed molecules, based on the predicted likelihood that that proposed molecule has the desired property and the approximate degree to which that proposed molecule is similar to the molecules represented in the dataset of training molecules.” [These elements are mental processes that can be performed by observation, evaluation, judgment, and opinion.]
Dependent Claim 16
“scoring each respective one of the proposed molecules based on the predicted likelihood of that proposed molecule having the desired property and the approximate degree to which that proposed molecules is similar to the molecules represented in the dataset of training molecules” [These elements are mental processes that can be performed by observation, evaluation, judgment, and opinion. While the claim recites that the scoring is based on certain elements, the term “based on” does not require any specific technical methodology that is distinguished from a mental process.]
“storing the data associated with that proposed molecule in the dataset of proposed molecules based on the score for that proposed molecule.” [These elements are considered to be additional elements besides the abstract idea. However, these elements are merely “adding insignificant extra-solution activity to the judicial exception” (MPEP § 2106.05(g)) since they merely amount to necessary data gathering or outputting, which identifies is identified in MPEP § 2106.05(g) as a form of extra-solution activity. Furthermore, for purposes of Step 2B analysis, these elements are well‐understood, routine, and conventional because they are limitations of “storing and retrieving information in memory,” which MPEP § 2106.05(d)(II) identifies as well‐understood, routine, and conventional computer functions.]
Dependent Claim 17
“wherein the computer-based property predictor is a computer-based graph neural network (GNN), the energy model is a Deep Energy Estimator Network (“DEEN”), and the search engine executes Monte Carlo tree searching (MCTS).” [These elements are considered to be additional elements besides the abstract idea but they do no more than generally link the use of a judicial exception to a particular technological environment or field of use (MPEP § 2106.05(h)), namely the technology environments of GNN, DEEN, and MCTS.]
Dependent Claim 18
“The computer-based method of claim 17, wherein the desired property is inhibition of SARS-CoV-1.” [This element merely further defines the mental processes recited in the independent claim, and does not introduce any further additional elements besides the abstract idea.]
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
1. Claims 1-6, 8, 10-15, and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Stojevic et al. (US 2021/0081804 A1) (“Stojevic”) in view of Saremi et al., “Deep Energy Estimator Networks,” arXiv:1805.08306v1 [stat.ML] 21 May 2018 (“Saremi”) (cited in an IDS) and Elton et al., “Deep learning for molecular generation and optimization - a review of the state of the art,” arXiv:1903.04388v1 [cs.LG] 11 Mar 2019 (“Elton”).
As to claim 1, Stojevic teaches a computer system for designing molecules with a desired property, the computer system comprising:
a computer-based processor; [[0274]: “a suitably programmed processor and associated memory”] and
a computer-readable medium storing computer-readable instructions that, when executed by the computer-based processor, cause the computer-based processor to embody: [[0274]: “a computer readable medium having stored thereon a program for carrying out any of the methods described herein and/or for embodying any of the apparatus features described herein.”]
a computer-based search engine configured to search for proposed molecules; [[0011]: “The machine learning method or system may efficiently search through, sample or otherwise analyze the tensor network representations to identify candidate small drug-like molecules with required properties, such as the binding affinity with respect to a target protein.”]
a computer-based property predictor to predict a likelihood of the desired property being present in the proposed molecules; [[0010]: “The machine learning method or system may provide, as its output, tensor network representations of the molecular quantum states of small drug-like molecules to a predictive model.” The representations are predictions, as described in [0119]: “the tensorial space is a wavefunction space. The latent space may represent physically relevant wavefunctions and/or desired properties of candidate chemical compounds. The latent space may represent a relevant physical operator decomposed as a tensor network.” [0137]: “Controlled optimization (a tailored search) is performed in the tensorial space with respect to complex cost functions that quantify the desirable properties, which are to be maximized (or minimized, as appropriate) over the search space”; [0122]: “said cost function representing desirable properties of said candidate chemical compounds.” That is, the cost function corresponds to a likelihood of the desired property being present in the molecule being searched.]
a computer-based […] model to approximate a degree of similarity between the proposed molecules and […], [[0213]: “The specific requirement is guided in terms of similarity to a particular molecule, using standard chemistry comparison methods, e.g. Tanimoto similarity, or using the latent space of a trained model, similarity measures between tensor network descriptions of the outputs (e.g. norm of the overlap between the candidate and target molecules).” See also [0086]: “similarity with respect to a set of molecules one wants to find close analogues of”; [0351]: “Correlating the tensorial space with a latent space may comprise optimising over the latent space (with respect to outputs of predictive models, or similarity with respect to a set of target molecules—e.g. using the Tanimoto index, or other metric)—for example so that it represents desirable properties of said candidate chemical compounds.” See also [0123]: “similarity with respect to a set of target molecules—e.g. using the Tanimoto index, or other metric)—for example so that it represents desirable properties of said candidate chemical compounds. This can be done in the context of GAN or VAE generative models, with or without the aforementioned tensor network extensions).”]
wherein the search is guided by the property predictor's predictions and the […] model's approximations. [[0011]: “The machine learning method or system may efficiently search through, sample or otherwise analyze the tensor network representations to identify candidate small drug-like molecules with required properties, such as the binding affinity with respect to a target protein.” [0212]: “the predictive model is configured to predict whether a candidate small drug-like molecule is appropriate to a specific requirement”; [0213]: “The specific requirement is guided in terms of similarity to a particular molecule.”]
Stojevic does not teach:
the limitation that the model for the approximation is an “energy” model; and
the limitation of the similarity being specifically with “molecules known to possess the desired property” [Examiner’s Note: Stojevic generally teaches that the training molecules are known to possess the desired property (see [0197]: “the tensor network representations include a training data set of small drug-like molecules and a target, in which the small drug-like molecules are known to bind to the target.”), but does not explicitly teach the detail that the similarity is with such molecules).
However, Saremi teaches the limitation of an “energy” model [Abstract: “In addition, the resulting deep energy estimator network (DEEN) is designed as products of experts. We present the utility of DEEN in learning the energy, the score function, and in single-step denoising experiments for synthetic and high-dimensional data.”]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of Stojevic with the teachings of Saremi by implementing the model for the approximation to use a deep energy estimator network model as taught in Saremi. The motivation would have been to use a type of model that enables the learning of the energy of a given data distribution (Saremi, § 1, last paragraph: “In this work, with score matching as the foundation, we introduce a scalable and efficient algorithm to learn the energy of any data distribution in an inference-free hierarchical framework.”), which is a known problem in machine learning in general (see Saremi, § 1, paragraph 1), and is applicable to the problem of capturing statistically meaningful distribution in a training data set described in Stojevic (see Stojevic, claim 3).
Elton teaches he remaining limitation of similarity with “molecules known to possess the desired property” [Page 15, last paragraph: “Another option for generating molecules close to training molecules but not too close is to have a reward for being similar to the training data but not too similar… Sim(S, T) ∈ [0, 1] is similarity scoring function which computes fingerprint-based similarity between the two molecules… This type of reward can be particularly useful for generating focused libraries of molecules very similar to a single target molecule or a small set of “actives” which are known to bind to a receptor. Note that most generative models can be tweaked to generate molecules close to a given molecule.” That is, Elton teaches the use of a similarity score for known molecules in the training set. Furthermore, as discussed on page 15, right column, this technique is used for generative models such as GAN and VAE.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Elton by implementing the similarity in the approximation to be a similarity with molecules in a training set that are known to possess the desired property, since doing so would enable the generation of generating molecules close to training molecules but not too close, as suggested by Elton (see parts cited above).
As to claim 2, the combination of Stojevic, Saremi, and Elton teaches the computer system of claim 1, further comprising:
a dataset of training molecules in the computer-readable medium, [Stojevic, abstract: “A training dataset may be used to train the machine learning system, and the training dataset is a tensor network representation of the molecular quantum states of small drug-like molecules.”]
wherein the dataset of training molecules contains data regarding the molecules known to possess the desired property. [[0197]: “the tensor network representations include a training data set of small drug-like molecules and a target, in which the small drug-like molecules are known to bind to the target.”]
As to claim 3, the combination of Stojevic, Saremi, and Elton teaches the computer system of claim 2, wherein the computer-based property predictor has been trained, in a supervised manner, to predict the likelihood of the desirable property being present in the proposed molecules, using the dataset of training molecules. [Stojevic, [0009]: “A training dataset may be used to train the machine learning system, and the training dataset is a tensor network representation of the molecular quantum states of small drug-like molecules.” Stojevic, [0214]: “the machine learning system is supervised, semi-supervised or unsupervised.” The context in which “supervised” is used, which is to describe a machine learning system that is undergoing training, implies that the training is performed in a supervised manner.]
As to claim 4, the combination of Stojevic, Saremi, and Elton teaches the computer system of claim 3, as set forth above.
Saremi further teaches “wherein the computer-based energy model has been trained, in an unsupervised manner, to model a distribution of data in the dataset of training molecules.” [Saremi, § 1, paragraph 1: “Learning the probability density of complex high-dimensional data is a challenging problem in machine learning.” This learning (training of the model) is with unsupervised learning. See, e.g., Saremi, § 5, paragraph 2: “Our approach fits within the general movement, prominent in recent years, which employs the SGD’s “unreasonable effectiveness” in training deep neural networks at the service of unsupervised learning…” Saremi, § 5, paragraph 3: “In our work here, we looked at another important problem in unsupervised learning, density estimation of unnormalized densities.” See also FIG. 1 and caption, which teaches “1000 SGD iterations in learning the energy,” which models the distribution of the training data.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have further combined the teachings of the references combined thus far, including the above further teachings of Saremi by implementing an unsupervised training for the energy model so as to arrive at the limitations of the instant dependent claim. The motivation for doing so is the same as the motivation given for the teachings of Saremi in the rejection of the parent independent claim.
As to claim 5, the combination of Stojevic, Saremi, and Elton teaches the computer-system of claim 4, wherein, for each molecule that the computer-based search engine proposes:
the computer-based property predictor predicts a likelihood that the proposed molecule has the desired property; [Stojevic, [0137]: “Controlled optimization (a tailored search) is performed in the tensorial space with respect to complex cost functions that quantify the desirable properties, which are to be maximized (or minimized, as appropriate) over the search space”; Stojevic, [0122]: “said cost function representing desirable properties of said candidate chemical compounds.” That is, the cost function corresponds to a likelihood of the desired property being present in the molecule being searched.] and
the computer-based energy model approximates a degree to which the proposed molecule is similar to molecules represented in the dataset of training molecules. [This is taught by the combination of references because Elton teaches the use of a similarity score specifically in the use case of known molecules in the training set, while Stojevic, [0197] teaches that the small drug-like molecules in the training data set are known to bind to the target. Therefore, the references combined thus far renders obvious the instant limitation under the same rationale given in the rejection of the parent independent claim.]
As to claim 6, the combination of Stojevic, Saremi, and Elton teaches the computer system of claim 5, further comprising:
a dataset of proposed molecules in the computer-readable medium, [Stojevic, [0118]: “outputting a further dataset, the further dataset comprising a filtered dataset and/or a set of novel chemical compounds (i.e. those not in the original dataset) of candidate chemical compounds.”]
wherein, for each molecule proposed by the search engine, the system determines whether to store data associated with that proposed molecule in the dataset of proposed molecules, based on the predicted likelihood that that proposed molecule has the desired property, and based on the approximate degree to which that proposed molecule is similar to the molecules represented in the dataset of training molecules. [Stojevic, [0118]: “outputting a further dataset, the further dataset comprising a filtered dataset and/or a set of novel chemical compounds (i.e. those not in the original dataset) of candidate chemical compounds.” See also Stojevic, [0120]: “The further dataset of candidate chemical compounds may be a filtered dataset of candidate chemical compounds”; Stojevic, [0085]: “The iterative cycle is repeated to maximise the quality of proposed molecules”; Stojevic, [0149]: “The output data may alternatively or additionally be a filtered version of the input data, corresponding to a smaller number of data points.” That is, the method is repeated such that the candidate compounds are repeatedly filtered. In regards to the limitation of “store data,” since the instant claim does not require any specific method of storage, the generation of the filtered set, which exists in the computer system and is thus stored in a memory of the system, reads on the limitation of storing data.]
As to claim 8, the combination of Stojevic, Saremi, and Elton teaches the computer system of claim 6, wherein the computer-based property predictor is a computer-based graph neural network (GNN), [Stojevic, [0110]-[0111]: “The manner in which tensor networks are incorporated into the GTN system can be classified as follows:… Using physics inspired tensor network decompositions within standard machine learning networks, for example to decompose RNNs (recurrent neural networks), convolutional neural networks, graph convolutional neural networks…”] […] and the search engine executes Monte Carlo tree searching (MCTS). [[0125]: “The generative model may comprise an (artificial) neural network, which may be in the form of e.g. a generative auto encoder, RNN, GAN, Monte-Carlo tree search model”]
Saremi further teaches “the energy model is a Deep Energy Estimator Network (“DEEN”)” [As discussed in the rejection of claim 1]. The motivation for incorporating the teachings of Saremi as given in the rejection of claim 1 also applies to the instant limitation.
As to claim 10, Stojevic teaches a computer-based method for designing molecules having a desired property, the computer-based method comprising:
providing a computer system comprising: a computer-based processor; and a computer-readable medium storing computer-readable instructions that, when executed by the computer-based processor, [[0128]: “a system for analysing a chemical compound dataset so as to determine suitable candidate chemical compounds, the system comprising: means for receiving a chemical compound dataset; a processor for processing the chemical compound dataset to determine a tensorial space for said chemical compound dataset; a processor for correlating said tensorial space with a latent space; and means for outputting a further dataset of candidate chemical compounds.” [0274]: “a computer readable medium having stored thereon a program for carrying out any of the methods described herein and/or for embodying any of the apparatus features described herein”] cause the computer-based processor to embody:
a computer-based property predictor to predict a likelihood that each respective one of a plurality of proposed molecules has the desired property; [[0010]: “The machine learning method or system may provide, as its output, tensor network representations of the molecular quantum states of small drug-like molecules to a predictive model.” The representations are predictions, as described in [0119]: “the tensorial space is a wavefunction space. The latent space may represent physically relevant wavefunctions and/or desired properties of candidate chemical compounds. The latent space may represent a relevant physical operator decomposed as a tensor network.” [0137]: “Controlled optimization (a tailored search) is performed in the tensorial space with respect to complex cost functions that quantify the desirable properties, which are to be maximized (or minimized, as appropriate) over the search space”; [0122]: “said cost function representing desirable properties of said candidate chemical compounds.” That is, the cost function corresponds to a likelihood of the desired property being present in the molecule being searched.]
a computer-based […] model to approximate a statistical similarity between each respective one of the proposed molecules and [[0213]: “The specific requirement is guided in terms of similarity to a particular molecule, using standard chemistry comparison methods, e.g. Tanimoto similarity, or using the latent space of a trained model, similarity measures between tensor network descriptions of the outputs (e.g. norm of the overlap between the candidate and target molecules).” See also [0086]: “similarity with respect to a set of molecules one wants to find close analogues of”; [0351]: “Correlating the tensorial space with a latent space may comprise optimising over the latent space (with respect to outputs of predictive models, or similarity with respect to a set of target molecules—e.g. using the Tanimoto index, or other metric)—for example so that it represents desirable properties of said candidate chemical compounds.” See also [0123]: “similarity with respect to a set of target molecules—e.g. using the Tanimoto index, or other metric)—for example so that it represents desirable properties of said candidate chemical compounds. This can be done in the context of GAN or VAE generative models, with or without the aforementioned tensor network extensions).”] […]; and
a computer-based search engine to conduct a search through a space of candidate molecules to propose molecules likely to have the desired property, wherein the search is guided by the property predictor's predictions and by the energy model's approximations. [[0011]: “The machine learning method or system may efficiently search through, sample or otherwise analyze the tensor network representations to identify candidate small drug-like molecules with required properties, such as the binding affinity with respect to a target protein.” [0212]: “the predictive model is configured to predict whether a candidate small drug-like molecule is appropriate to a specific requirement”; [0213]: “The specific requirement is guided in terms of similarity to a particular molecule.”]
Stojevic does not teach:
the limitation that the model for the approximation is an “energy” model; and
the limitation of the similarity being specifically with “a set of molecules represented in a stored dataset of training molecules, wherein the dataset of training molecules contains data regarding a set of molecules known to possess the desired property.” [Examiner’s Note: Stojevic generally teaches that the training molecules are known to possess the desired property (see [0197]: “the tensor network representations include a training data set of small drug-like molecules and a target, in which the small drug-like molecules are known to bind to the target.”), but does not explicitly teach the detail that the similarity is with such molecules).]
However, Saremi teaches the limitation of an “energy” model [Abstract: “In addition, the resulting deep energy estimator network (DEEN) is designed as products of experts. We present the utility of DEEN in learning the energy, the score function, and in single-step denoising experiments for synthetic and high-dimensional data.”]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of Stojevic with the teachings of Saremi by implementing the model for the approximation to use a deep energy estimator network model as taught in Saremi. The motivation would have been to use a type of model that enables the learning of the energy of a given data distribution (Saremi, § 1, last paragraph: “In this work, with score matching as the foundation, we introduce a scalable and efficient algorithm to learn the energy of any data distribution in an inference-free hierarchical framework.”), which is a known problem in machine learning in general (see Saremi, § 1, paragraph 1), and is applicable to the problem of capturing statistically meaningful distribution in a training data set described in Stojevic (see Stojevic, claim 3).
Elton teaches he remaining limitation of similarity with “a set of molecules represented in a stored dataset of training molecules, wherein the dataset of training molecules contains data regarding a set of molecules known to possess the desired property” [Page 15, last paragraph: “Another option for generating molecules close to training molecules but not too close is to have a reward for being similar to the training data but not too similar… Sim(S, T) ∈ [0, 1] is similarity scoring function which computes fingerprint-based similarity between the two molecules… This type of reward can be particularly useful for generating focused libraries of molecules very similar to a single target molecule or a small set of “actives” which are known to bind to a receptor. Note that most generative models can be tweaked to generate molecules close to a given molecule.” That is, Elton teaches the use of a similarity score for known molecules in the training set. Furthermore, as discussed on page 15, right column, this technique is used for generative models such as GAN and VAE.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Elton by implementing the similarity in the approximation to be a similarity with “a set of molecules represented in a stored dataset of training molecules, wherein the dataset of training molecules contains data regarding a set of molecules known to possess the desired property,” since doing so would enable the generation of generating molecules close to training molecules but not too close, as suggested by Elton (see parts cited above).
As to claim 11, the combination of Stojevic, Saremi, and Elton teaches the computer-based method of claim 10, further comprising:
storing the dataset of training molecules in the computer-readable medium; [Stojevic, [0009]: “A training dataset may be used to train the machine learning system, and the training dataset is a tensor network representation of the molecular quantum states of small drug-like molecules.” Since the training dataset is used in a machine learning operation, it is implied that it is stored in a computer-readable medium.] and
training, in a supervised manner, the computer-based property predictor to predict a likelihood that the desirable property is present in a molecule using the dataset of training molecules. [Stojevic, [0009]: “A training dataset may be used to train the machine learning system, and the training dataset is a tensor network representation of the molecular quantum states of small drug-like molecules.” Stojevic, [0214]: “the machine learning system is supervised, semi-supervised or unsupervised.” The context in which “supervised” is used, which is to describe a machine learning system that is undergoing training, implies that the training is performed in a supervised manner.]
As to claim 12, the combination of Stojevic, Saremi, and Elton teaches the computer-based method of claim 11, further comprising:
Saremi further teaches “training, in an unsupervised manner, the energy model to model a distribution of data regarding the training molecules in the dataset of training molecules.” [Saremi, § 1, paragraph 1: “Learning the probability density of complex high-dimensional data is a challenging problem in machine learning.” This learning (training of the model) is with unsupervised learning. See, e.g., Saremi, § 5, paragraph 2: “Our approach fits within the general movement, prominent in recent years, which employs the SGD’s “unreasonable effectiveness” in training deep neural networks at the service of unsupervised learning…” Saremi, § 5, paragraph 3: “In our work here, we looked at another important problem in unsupervised learning, density estimation of unnormalized densities.” See also FIG. 1 and caption, which teaches “1000 SGD iterations in learning the energy,” which models the distribution of the training data.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have further combined the teachings of the references combined thus far, including the above further teachings of Saremi by implementing an unsupervised training for the energy model so as to arrive at the limitations of the instant dependent claim. The motivation for doing so is the same as the motivation given for the teachings of Saremi in the rejection of the parent independent claim.
As to claim 13, the combination of Stojevic, Saremi, and Elton teaches the computer-based method of claim 12, further comprising:
conducting the search through the space of candidate molecules with the computer-based search engine guided by the property predictor's predictions and the energy model's approximations. [Stojevic, [0011]: “The machine learning method or system may efficiently search through, sample or otherwise analyze the tensor network representations to identify candidate small drug-like molecules with required properties, such as the binding affinity with respect to a target protein.” Stojevic, [0212]: “the predictive model is configured to predict whether a candidate small drug-like molecule is appropriate to a specific requirement”; Stojevic, [0213]: “The specific requirement is guided in terms of similarity to a particular molecule.”]
As to claim 14, the combination of Stojevic, Saremi, and Elton teaches the computer-based method of claim 13, further comprising, for each molecule that the computer-based search engine proposes:
predicting, with the computer-based property predictor, a likelihood that that proposed molecule has the desired property; [Stojevic, [0010]: “The machine learning method or system may provide, as its output, tensor network representations of the molecular quantum states of small drug-like molecules to a predictive model.” The representations are predictions, as described in Stojevic, [0119]: “the tensorial space is a wavefunction space. The latent space may represent physically relevant wavefunctions and/or desired properties of candidate chemical compounds. The latent space may represent a relevant physical operator decomposed as a tensor network.” Stojevic, [0137]: “Controlled optimization (a tailored search) is performed in the tensorial space with respect to complex cost functions that quantify the desirable properties, which are to be maximized (or minimized, as appropriate) over the search space”; Stojevic, [0122]: “said cost function representing desirable properties of said candidate chemical compounds.” That is, the cost function corresponds to a likelihood of the desired property being present in the molecule being searched.] and
approximating, with the computer-based energy model, a degree to which that proposed molecule is similar to molecules represented in the dataset of training molecules. [This limitation is taught by the prior art for for the reasons given in the rejection of the parent independent claim for the corresponding limitations of the energy model in the parent independent claim.]
As to claim 15, the combination of Stojevic, Saremi, and Elton teaches the computer-based method of claim 14, further comprising:
storing, in the computer-readable medium, a dataset of proposed molecules; [Stojevic, [0118]: “outputting a further dataset, the further dataset comprising a filtered dataset and/or a set of novel